Comparing Topology based Collective Communication Algorithms

1
Comparing Topology based Collective Communication
Algorithms

• Vishal Sharda
• Ashima Gupta

2
Work progress

• No up-to-date open source solution available for
  fully mapping the network to build upon
• Nomad, Cheops, Nagios, ENV, Argus etc. outdated
• Use of a third-party service to determine the
  topology
• Java applet to map SERC network

3
Status of current study

• Nearly all the research on this problem till now
  considers tightly-coupled systems.
• Thus, most of the algorithms assume
  point-to-point connectivity.
• So, these have to be modified to adapt to the
  network of heterogeneous workstations.

4
All to all broadcast

• Also known as multinode broadcast
• Generalization of one-to-all broadcast in which
  all the processors simultaneously initiate a
  broadcast.
• Different processors may send out different
  message.

5
Existing Algorithms

• Direct exchange
• Circular all-to-all
• E1 algorithm
• Liquid Schedule
• Algorithms for specific topologies like start and
  mesh.

6
Direct Exchange

• Simplest approach
• Assumes point to point connectivity
• Each node simultaneously sends data to each other
  node
• Involves lot of congestion

7
Circular all-to-all

• Let p be the no. of nodes
• For each node i
• for step k in 1..p, node i
• sends to (ik)mod p
• receives from(i-kP)mod p

8
E1 algorithm

• One node receives messages from all other
  nodes(becomes an expert)
• Experts are formed by recursively doubling
  existing experts.

9
Liquid schedule algorithm

• Traffic is the set of all collective exchanges.
• A simultaneous sub traffic is part of the traffic
  involing non-congesting transfers.
• Identify the bottleneck links in the network.
• A liquid schedule is such that all the bottleneck
  links are utilized in every sub-traffic.

10
Need for Simulation

• Deciding the packet to follow a particular route
  needs to bypass router decisions
• Needs DLL operations.
• Focus on comparing the algorithms
• Incorporate simulation-based study like standard
  experiments.

11
Basis of Simulation

• Execute and compare the algorithms for the
  network specifed at the interface.
• Some algorithms will take bandwidth of the links
  into account.
• Depending on the input, an algorithm may or may
  not show good results.

12
Implementation Approach

• Network represented as weighted undirected graph
  with weights inversely proportional to bandwidth.
• Simulating n processes on different nodes in a
  network with Java threads.
• First of all, each node will compute the shortest
  path to all other nodes using single source
  shortest path algorithm.

13
Handling Collision

• Lot of collision involved in all-to-all
  broadcast.
• Several approaches possible like partitioning
  into subnets , choosing alternate link.
• Our approach will be to stick to the shortest
  link and if collision detected then wait for
  random time and sense again.

14
References

• Papers
• "ECO Efficient Collective Operations for
  Communication on Heterogeneous Networks",Bruce B.
  Lowekamp and Adam Begueliny.
• "Network Topology Aware Scheduling of Collective
  Communications",Emin Gabrielyan, Roger D. Hersch.
• "On General Results for all-to-all broadcast",
  Ming-Syan Chen et. al
• Efficient all-to-all broadcast in star graph
  interconnecion networks",Yu-Chee Tseng et.al
• Websites
• freemap.qualys.com